基于尺度分割和特征优选的外破检测技术

External Break Detection Technology Based on Scale Segmentation and Feature Optimization

高分辨率遥感卫星是识别外力破坏的巡查中不可或缺的有效工具,而实现外力破坏迹象识别的关键在于变化检测技术的应用。文章提出一种基于最优尺度分割和特征优选的对象级变化检测方法,首先通过尺度评估参数确定最优分割尺度;然后基于最优分割尺度结果,分别采用J-M距离算法和XGBoost模型对多源特征集进行优选,对比分析筛选出合适的特征集合;最后采用支持向量机(SVM)分类器计算各特征参数间的欧式距离并进行二分类,得到变化和未变化两种分类结果。结果表明,XGBoost模型算法进行特征优选后的精度指标均高于J-M距离算法的变化检测结果,并将同一分割尺度和特征集合应用于不同分辨率影像中验证变化检测结果,在面积和对象两种精度评估中,F1分数分别达到了83.31和84.09,而查准率均超过了79%,查全率也均超过了86%,说明文章使用的算法在不同分辨率下能获得较高的检测精度,同时减少了人工对特征选择和特征集建立规则的干预,为电缆沿线变化检测提供有效的技术支撑。

High resolution remote sensing satellite is an indispensable and effective tool in the detection of external damage,and the key to realize the identification of external damage signs is the application of change detection technology.In this paper,an object-level change detection method based on optimal scale segmentation and feature optimization is proposed.Firstly,the optimal segmentation scale is determined by scale evaluation parameters.Then,based on the results of the optimal segmentation scale,J-M distance algorithm and XGBoost model were respectively used to optimize the multi-source feature sets,and the appropriate feature sets were screened by comparative analysis.Finally,a support vector machine(SVM)classifier was used to calculate the Euclidean distance between the feature parameters and classify them,and two classification results were obtained,namely,changed and unchanged.The results show that the accuracy indexes of XGBoost model algorithm after feature optimization are all higher than the change detection results of J-M distance algorithm,and the same segmentation scale and feature set are applied to validate the change detection results in images with different resolutions,in the accuracy evaluation based on area and object,F1 scores reached 83.31%and 84.09% respectively.Meanwhile the accuracy rate exceeded 79%and the recall rate exceeded 86%,which indicates that the algorithm used in this paper can obtain higher detection accuracy under different resolutions,and the manual intervention in feature selection and feature set establishment rules is reduced,which provides effective technical support for change detection along the cable.